[PATCH v3] perf: Fix use-after-free when perf mmap() revival races with the last munmap()

[PATCH v3] perf: Fix use-after-free when perf mmap() revival races with the last munmap()

[Yilin Zhang]

perf_mmap_close() drops rb->mmap_count *without* holding
event->mmap_mutex (the refcount_dec_and_test() right before the
refcount_dec_and_mutex_lock() of event->mmap_count). A concurrent
perf_mmap_rb() can slot its entire "revival" path into that window
(perf_mmap holds event->mmap_mutex for its whole duration, including
rb_alloc):

  munmap side (perf_mmap_close)          mmap side (perf_mmap_rb)
  -----------------------------------    --------------------------------
  rb->mmap_count 1 -> 0   (no lock)      (holds event->mmap_mutex)
                                         inc_not_zero(rb->mmap_count) fails
                                         ring_buffer_attach(event, NULL)
                                         rb_alloc() + attach new rb
                                         refcount_set(&event->mmap_count, 1)
  lock; event->mmap_count 1 -> 0
  ring_buffer_attach(event, NULL)
  ring_buffer_put() -> frees the *new* rb

The revival's refcount_set(&event->mmap_count, 1) is an invisible
1 -> 1 write: the close frees the just-revived buffer although the
other process still has it mapped -- a page-level use-after-free
allowing local privilege escalation to root by any unprivileged user
(default kernel.perf_event_paranoid=2).

Swap the order of the two counter updates: event->mmap_count is
dropped first via refcount_dec_and_mutex_lock(), so its 1 -> 0
transition and the ring_buffer_attach() stay serialized with
perf_mmap(). rb->mmap_count == 0 then implies every event using the
buffer is detached already, so the result of the rb->mmap_count drop
can gate the remaining teardown directly and detach_rest is no longer
needed.

An earlier fix for this race from Kyle Zeng and David Lee takes
event->mmap_mutex around both counter updates [0]; here the not-last
close stays lockless.

Fixes: 59741451b49c ("perf: Identify the 0->1 transition for event::mmap_count")
Link: https://lore.kernel.org/linux-perf-users/20260804060931.711308-1-david.lee@trailofbits.com/ [0]
Cc: stable@vger.kernel.org # 6.18+
Reported-by: Kimi Security Team <bug-report@moonshot.ai>
Tested-by: Weiming Shi <shiweiming@moonshot.ai>
Signed-off-by: Yilin Zhang <yilinzhang@moonshot.ai>
---
v2: https://lore.kernel.org/linux-perf-users/91825d0335d2f7cc017ee23886ffeef44c1cc633.f061f277.0045.493a.937e.53242f19a7de@feishu.cn/T/#u
v3: swap the counter drop order instead of serializing both under
    mmap_mutex; keeps the lockless fast path, drops detach_rest, and
    references the earlier independent fix [0]

 kernel/events/core.c | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/kernel/events/core.c b/kernel/events/core.c
index a6c8e38a3110..f56f9d9e4f01 100644
--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -7029,7 +7029,6 @@ static void perf_mmap_close(struct vm_area_struct *vma)
 	mapped_f unmapped = get_mapped(event, event_unmapped);
 	struct perf_buffer *rb = ring_buffer_get(event);
 	struct user_struct *mmap_user = rb->mmap_user;
-	bool detach_rest = false;
 
 	/* FIXIES vs perf_pmu_unregister() */
 	if (unmapped)
@@ -7060,17 +7059,14 @@ static void perf_mmap_close(struct vm_area_struct *vma)
 		mutex_unlock(&rb->aux_mutex);
 	}
 
-	if (refcount_dec_and_test(&rb->mmap_count))
-		detach_rest = true;
-
-	if (!refcount_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
-		goto out_put;
-
-	ring_buffer_attach(event, NULL);
-	mutex_unlock(&event->mmap_mutex);
+	if (refcount_dec_and_mutex_lock(&event->mmap_count,
+					&event->mmap_mutex)) {
+		ring_buffer_attach(event, NULL);
+		mutex_unlock(&event->mmap_mutex);
+	}
 
 	/* If there's still other mmap()s of this buffer, we're done. */
-	if (!detach_rest)
+	if (!refcount_dec_and_test(&rb->mmap_count))
 		goto out_put;
 
 	/*
-- 
2.43.0

Re: [PATCH v3] perf: Fix use-after-free when perf mmap() revival races with the last munmap()

[Peter Zijlstra]

On Mon, Aug 31, 2026 at 09:31:52PM +0800, Yilin Zhang wrote:
> perf_mmap_close() drops rb->mmap_count *without* holding
> event->mmap_mutex (the refcount_dec_and_test() right before the
> refcount_dec_and_mutex_lock() of event->mmap_count). A concurrent
> perf_mmap_rb() can slot its entire "revival" path into that window
> (perf_mmap holds event->mmap_mutex for its whole duration, including
> rb_alloc):
> 
>   munmap side (perf_mmap_close)          mmap side (perf_mmap_rb)
>   -----------------------------------    --------------------------------
>   rb->mmap_count 1 -> 0   (no lock)      (holds event->mmap_mutex)
>                                          inc_not_zero(rb->mmap_count) fails
>                                          ring_buffer_attach(event, NULL)
>                                          rb_alloc() + attach new rb
>                                          refcount_set(&event->mmap_count, 1)
>   lock; event->mmap_count 1 -> 0
>   ring_buffer_attach(event, NULL)
>   ring_buffer_put() -> frees the *new* rb
> 
> The revival's refcount_set(&event->mmap_count, 1) is an invisible
> 1 -> 1 write: the close frees the just-revived buffer although the
> other process still has it mapped -- a page-level use-after-free
> allowing local privilege escalation to root by any unprivileged user
> (default kernel.perf_event_paranoid=2).
> 
> Swap the order of the two counter updates: event->mmap_count is
> dropped first via refcount_dec_and_mutex_lock(), so its 1 -> 0
> transition and the ring_buffer_attach() stay serialized with
> perf_mmap(). rb->mmap_count == 0 then implies every event using the
> buffer is detached already, so the result of the rb->mmap_count drop
> can gate the remaining teardown directly and detach_rest is no longer
> needed.
> 
> An earlier fix for this race from Kyle Zeng and David Lee takes
> event->mmap_mutex around both counter updates [0]; here the not-last
> close stays lockless.
> 
> Fixes: 59741451b49c ("perf: Identify the 0->1 transition for event::mmap_count")
> Link: https://lore.kernel.org/linux-perf-users/20260804060931.711308-1-david.lee@trailofbits.com/ [0]
> Cc: stable@vger.kernel.org # 6.18+
> Reported-by: Kimi Security Team <bug-report@moonshot.ai>
> Tested-by: Weiming Shi <shiweiming@moonshot.ai>
> Signed-off-by: Yilin Zhang <yilinzhang@moonshot.ai>
> ---
> v2: https://lore.kernel.org/linux-perf-users/91825d0335d2f7cc017ee23886ffeef44c1cc633.f061f277.0045.493a.937e.53242f19a7de@feishu.cn/T/#u
> v3: swap the counter drop order instead of serializing both under
>     mmap_mutex; keeps the lockless fast path, drops detach_rest, and
>     references the earlier independent fix [0]

I presume this means you and your AI agree with my pre-wakeup-juice
morning musings?

Also, you seem to have lost the 'helpful' comments that I drafted :-(


>  kernel/events/core.c | 16 ++++++----------
>  1 file changed, 6 insertions(+), 10 deletions(-)
> 
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index a6c8e38a3110..f56f9d9e4f01 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -7029,7 +7029,6 @@ static void perf_mmap_close(struct vm_area_struct *vma)
>  	mapped_f unmapped = get_mapped(event, event_unmapped);
>  	struct perf_buffer *rb = ring_buffer_get(event);
>  	struct user_struct *mmap_user = rb->mmap_user;
> -	bool detach_rest = false;
>  
>  	/* FIXIES vs perf_pmu_unregister() */
>  	if (unmapped)
> @@ -7060,17 +7059,14 @@ static void perf_mmap_close(struct vm_area_struct *vma)
>  		mutex_unlock(&rb->aux_mutex);
>  	}
>  
> -	if (refcount_dec_and_test(&rb->mmap_count))
> -		detach_rest = true;
> -
> -	if (!refcount_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
> -		goto out_put;
> -
> -	ring_buffer_attach(event, NULL);
> -	mutex_unlock(&event->mmap_mutex);
> +	if (refcount_dec_and_mutex_lock(&event->mmap_count,
> +					&event->mmap_mutex)) {
> +		ring_buffer_attach(event, NULL);
> +		mutex_unlock(&event->mmap_mutex);
> +	}
>  
>  	/* If there's still other mmap()s of this buffer, we're done. */
> -	if (!detach_rest)
> +	if (!refcount_dec_and_test(&rb->mmap_count))
>  		goto out_put;
>  
>  	/*
> -- 
> 2.43.0

Re: [PATCH v3] perf: Fix use-after-free when perf mmap() revival races with the last munmap()

[Yilin Zhang]

Sorry for just posting v3 instead of replying to you here first. It
is your ordering, with one refinement.

In your version the lockless refcount_dec(&rb->mmap_count) can be
the one to drop it to 0: dec_not_one() drops event->mmap_count
without the mutex, so a second closer can zero the event count in
between.

That warns, and detach_rest is never set, so the last mapping skips
the detach from redirecting events. On 8 vCPU QEMU it hits within a
minute:

	refcount_t: decrement hit 0; leaking memory.
	RIP: perf_mmap_close+0x221  (refcount_warn_saturate(REFCOUNT_DEC_LEAK))

The rb drop itself now gates the teardown: rb->mmap_count == 0
implies every event is detached already.

> Also, you seem to have lost the 'helpful' comments that I drafted :-(

Sorry for dropping your comments and the Suggested-by tag. v4 restores
both and follows shortly.

Re: [PATCH v3] perf: Fix use-after-free when perf mmap() revival races with the last munmap()

[sashiko-bot]

Thank you for your contribution! Sashiko AI review found 1 potential issue(s) to consider:

Pre-existing issues:
- [High] Permanent leak of locked_vm accounting when mmapping the same perf event ring buffer or AUX buffer multiple times.
--

commit bd560c15b9fb2009d96bf58b9046c04989905b44
Author: Yilin Zhang <yilinzhang@moonshot.ai>

perf: Fix use-after-free when perf mmap() revival races with the last munmap()

This commit fixes a use-after-free vulnerability in the perf events subsystem
by swapping the order of counter decrements in perf_mmap_close(). The change
ensures that event->mmap_count is decremented first under the mmap_mutex,
closing a race window where a newly attached ring buffer could be mistakenly
freed.

Link: https://lore.kernel.org/linux-perf-users/20260804060931.711308-1-david.lee@trailofbits.com/

> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index a6c8e38a31104..f56f9d9e4f01b 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -7060,17 +7059,14 @@ static void perf_mmap_close(struct vm_area_struct *vma)
>  		mutex_unlock(&rb->aux_mutex);
>  	}
>  
> -	if (refcount_dec_and_test(&rb->mmap_count))
> -		detach_rest = true;
> -
> -	if (!refcount_dec_and_mutex_lock(&event->mmap_count, &event->mmap_mutex))
> -		goto out_put;
> -
> -	ring_buffer_attach(event, NULL);
> -	mutex_unlock(&event->mmap_mutex);
> +	if (refcount_dec_and_mutex_lock(&event->mmap_count,
> +					&event->mmap_mutex)) {
> +		ring_buffer_attach(event, NULL);
> +		mutex_unlock(&event->mmap_mutex);
> +	}
>  
>  	/* If there's still other mmap()s of this buffer, we're done. */
> -	if (!detach_rest)
> +	if (!refcount_dec_and_test(&rb->mmap_count))
>  		goto out_put;

[Severity: High]
This is a pre-existing issue, but does the early return here leak the
locked_vm accounting when the same perf event ring buffer or AUX buffer is
mmapped multiple times?

Looking at the mmap side in perf_mmap_rb() and perf_mmap_aux(), secondary 
mmaps repeatedly charge the user's locked_vm:

kernel/events/core.c:perf_mmap_rb() {
	...
	perf_mmap_account(vma, user_extra, extra);
	refcount_inc(&event->mmap_count);
	...
}

kernel/events/core.c:perf_mmap_aux() {
	...
	perf_mmap_account(vma, user_extra, extra);
	refcount_inc(&event->mmap_count);
	...
}

However, the unaccounting in perf_mmap_close() only happens exactly once when
the last mmap is closed and rb->mmap_count reaches zero:

kernel/events/core.c:perf_mmap_close() {
	...
	if (!refcount_dec_and_test(&rb->mmap_count))
		goto out_put;
	...
	perf_mmap_unaccount(vma, rb);
out_put:
	ring_buffer_put(rb);
}

Because locked_vm is used system-wide to enforce RLIMIT_MEMLOCK (e.g., by
io_uring, BPF, XDP, and secretmem), an attacker could trivially loop calling
mmap() on the same file descriptor to artificially inflate this counter until
the limit is exhausted.

Does this create a persistent Denial of Service for all processes running
under that UID, making the system fully unusable for the affected user's
services until all processes of that user exit?

-- 
Sashiko AI review · https://sashiko.dev/#/patchset/20260831133152.1231045-1-yilinzhang@moonshot.ai?part=1